Shanzhe Yang

LLMs can get a 27.8% boost in mathematical reasoning by fusing a hardware-efficient optimal control layer directly into their architecture, enabling planning before prediction.